Electronic Thesis and Dissertation Repository


Master of Science


Computer Science


Dr. Olga Veksler


Symmetry is a useful segmentation cue. We develop an algorithm for segmenting a single symmetric object from the background. Our algorithm is formulated in the principled global optimization framework. Thus we can incorporate all the useful segmentation cues in the global energy function, in addition to the symmetry shape prior. We use the standard cues of regular boundary and coherent object (background) appearance. Our algorithm consists of two stages. The first stage, based on seam carving, detects a set of symmetry axis candidates. Symmetry axis is detected by first finding image “seams” that are aligned with intensity gradients and then matching them based on pairwise symmetry. The second stage formulates symmetric object segmentation in discrete optimization framework. We choose the longest symmetry axis as the object axis. Object symmetry is encouraged through submodular long-range pairwise terms. These pairwise terms are submodular, so optimization with a graph cut is applicable. We demonstrate the effectiveness of symmetry cue on a new symmetric object dataset.